fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[CPU/Backend] Refactor shared backend logic into CRTP base

Browse Source
This commit is contained in:
Herman S.
2026-03-17 09:22:10 +09:00
parent b621e89c0b
commit b8912a4c79
4 changed files with 407 additions and 423 deletions
Download Patch File Download Diff File Expand all files Collapse all files

323
src/xenia/cpu/backend/x64/x64_code_cache.cc
View File

@@ -9,7 +9,6 @@
#include "xenia/cpu/backend/x64/x64_code_cache.h"
#include <cstdlib>
#include <cstring>
#if ENABLE_VTUNE
@@ -17,254 +16,36 @@
#pragma comment(lib, "../third_party/vtune/lib64/jitprofiling.lib")
#endif
#if ENABLE_VTUNE
#include "third_party/fmt/include/fmt/format.h"
#include "xenia/base/assert.h"
#include "xenia/base/clock.h"
#include "xenia/base/literals.h"
#include "xenia/base/logging.h"
#include "xenia/base/math.h"
#include "xenia/base/memory.h"
#include "xenia/cpu/function.h"
#include "xenia/cpu/module.h"
#endif
namespace xe {
namespace cpu {
namespace backend {
namespace x64 {
using namespace xe::literals;
bool X64CodeCache::Initialize() { return CodeCacheBase::Initialize(); }
X64CodeCache::X64CodeCache() = default;
X64CodeCache::~X64CodeCache() {
if (indirection_table_base_) {
xe::memory::DeallocFixed(indirection_table_base_, kIndirectionTableSize,
xe::memory::DeallocationType::kRelease);
}
// Unmap all views and close mapping.
if (mapping_ != xe::memory::kFileMappingHandleInvalid) {
if (generated_code_write_base_ &&
generated_code_write_base_ != generated_code_execute_base_) {
xe::memory::UnmapFileView(mapping_, generated_code_write_base_,
kGeneratedCodeSize);
}
if (generated_code_execute_base_) {
xe::memory::UnmapFileView(mapping_, generated_code_execute_base_,
kGeneratedCodeSize);
}
xe::memory::CloseFileMappingHandle(mapping_, file_name_);
mapping_ = xe::memory::kFileMappingHandleInvalid;
}
void X64CodeCache::FillCode(void* write_address, size_t size) {
std::memset(write_address, 0xCC, size);
}
bool X64CodeCache::Initialize() {
indirection_table_base_ = reinterpret_cast<uint8_t*>(xe::memory::AllocFixed(
reinterpret_cast<void*>(kIndirectionTableBase), kIndirectionTableSize,
xe::memory::AllocationType::kReserve,
xe::memory::PageAccess::kReadWrite));
if (!indirection_table_base_) {
XELOGE("Unable to allocate code cache indirection table");
XELOGE(
"This is likely because the {:X}-{:X} range is in use by some other "
"system DLL",
static_cast<uint64_t>(kIndirectionTableBase),
kIndirectionTableBase + kIndirectionTableSize);
}
// Create mmap file. This allows us to share the code cache with the debugger.
file_name_ = fmt::format("xenia_code_cache_{}", Clock::QueryHostTickCount());
mapping_ = xe::memory::CreateFileMappingHandle(
file_name_, kGeneratedCodeSize, xe::memory::PageAccess::kExecuteReadWrite,
false);
if (mapping_ == xe::memory::kFileMappingHandleInvalid) {
XELOGE("Unable to create code cache mmap");
return false;
}
// Map generated code region into the file. Pages are committed as required.
if (xe::memory::IsWritableExecutableMemoryPreferred()) {
generated_code_execute_base_ =
reinterpret_cast<uint8_t*>(xe::memory::MapFileView(
mapping_, reinterpret_cast<void*>(kGeneratedCodeExecuteBase),
kGeneratedCodeSize, xe::memory::PageAccess::kExecuteReadWrite, 0));
generated_code_write_base_ = generated_code_execute_base_;
if (!generated_code_execute_base_ || !generated_code_write_base_) {
XELOGE("Unable to allocate code cache generated code storage");
XELOGE(
"This is likely because the {:X}-{:X} range is in use by some other "
"system DLL",
uint64_t(kGeneratedCodeExecuteBase),
uint64_t(kGeneratedCodeExecuteBase + kGeneratedCodeSize));
return false;
}
} else {
generated_code_execute_base_ =
reinterpret_cast<uint8_t*>(xe::memory::MapFileView(
mapping_, reinterpret_cast<void*>(kGeneratedCodeExecuteBase),
kGeneratedCodeSize, xe::memory::PageAccess::kExecuteReadOnly, 0));
generated_code_write_base_ =
reinterpret_cast<uint8_t*>(xe::memory::MapFileView(
mapping_, reinterpret_cast<void*>(kGeneratedCodeWriteBase),
kGeneratedCodeSize, xe::memory::PageAccess::kReadWrite, 0));
if (!generated_code_execute_base_ || !generated_code_write_base_) {
XELOGE("Unable to allocate code cache generated code storage");
XELOGE(
"This is likely because the {:X}-{:X} and {:X}-{:X} ranges are in "
"use by some other system DLL",
uint64_t(kGeneratedCodeExecuteBase),
uint64_t(kGeneratedCodeExecuteBase + kGeneratedCodeSize),
uint64_t(kGeneratedCodeWriteBase),
uint64_t(kGeneratedCodeWriteBase + kGeneratedCodeSize));
return false;
}
}
// Preallocate the function map to a large, reasonable size.
generated_code_map_.reserve(kMaximumFunctionCount);
return true;
void X64CodeCache::FlushCodeRange(void* address, size_t size) {
// x86-64 has coherent I/D caches; no flush needed.
}
void X64CodeCache::set_indirection_default(uint32_t default_value) {
indirection_default_value_ = default_value;
}
void X64CodeCache::AddIndirection(uint32_t guest_address,
uint32_t host_address) {
if (!indirection_table_base_) {
return;
}
uint32_t* indirection_slot = reinterpret_cast<uint32_t*>(
indirection_table_base_ + (guest_address - kIndirectionTableBase));
*indirection_slot = host_address;
}
void X64CodeCache::CommitExecutableRange(uint32_t guest_low,
uint32_t guest_high) {
if (!indirection_table_base_) {
return;
}
// Commit the memory.
xe::memory::AllocFixed(
indirection_table_base_ + (guest_low - kIndirectionTableBase),
guest_high - guest_low, xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kReadWrite);
// Fill memory with the default value.
uint32_t* p = reinterpret_cast<uint32_t*>(indirection_table_base_);
for (uint32_t address = guest_low; address < guest_high; ++address) {
p[(address - kIndirectionTableBase) / 4] = indirection_default_value_;
}
}
void X64CodeCache::PlaceHostCode(uint32_t guest_address, void* machine_code,
const EmitFunctionInfo& func_info,
void*& code_execute_address_out,
void*& code_write_address_out) {
// Same for now. We may use different pools or whatnot later on, like when
// we only want to place guest code in a serialized cache on disk.
PlaceGuestCode(guest_address, machine_code, func_info, nullptr,
code_execute_address_out, code_write_address_out);
}
void X64CodeCache::PlaceGuestCode(uint32_t guest_address, void* machine_code,
const EmitFunctionInfo& func_info,
GuestFunction* function_info,
void*& code_execute_address_out,
void*& code_write_address_out) {
// Hold a lock while we bump the pointers up. This is important as the
// unwind table requires entries AND code to be sorted in order.
size_t low_mark;
size_t high_mark;
uint8_t* code_execute_address;
UnwindReservation unwind_reservation;
{
auto global_lock = global_critical_region_.Acquire();
low_mark = generated_code_offset_;
// Reserve code.
// Always move the code to land on 16b alignment.
code_execute_address =
generated_code_execute_base_ + generated_code_offset_;
code_execute_address_out = code_execute_address;
uint8_t* code_write_address =
generated_code_write_base_ + generated_code_offset_;
code_write_address_out = code_write_address;
generated_code_offset_ += xe::round_up(func_info.code_size.total, 16);
auto tail_write_address =
generated_code_write_base_ + generated_code_offset_;
// Reserve unwind info.
// We go on the high size of the unwind info as we don't know how big we
// need it, and a few extra bytes of padding isn't the worst thing.
unwind_reservation = RequestUnwindReservation(generated_code_write_base_ +
generated_code_offset_);
generated_code_offset_ += xe::round_up(unwind_reservation.data_size, 16);
auto end_write_address =
generated_code_write_base_ + generated_code_offset_;
high_mark = generated_code_offset_;
// Store in map. It is maintained in sorted order of host PC dependent on
// us also being append-only.
generated_code_map_.emplace_back(
(uint64_t(code_execute_address - generated_code_execute_base_) << 32) |
generated_code_offset_,
function_info);
// TODO(DrChat): The following code doesn't really need to be under the
// global lock except for PlaceCode (but it depends on the previous code
// already being ran)
// If we are going above the high water mark of committed memory, commit
// some more. It's ok if multiple threads do this, as redundant commits
// aren't harmful.
size_t old_commit_mark, new_commit_mark;
do {
old_commit_mark = generated_code_commit_mark_;
if (high_mark <= old_commit_mark) break;
new_commit_mark = old_commit_mark + 16_MiB;
if (generated_code_execute_base_ == generated_code_write_base_) {
xe::memory::AllocFixed(generated_code_execute_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kExecuteReadWrite);
} else {
xe::memory::AllocFixed(generated_code_execute_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kExecuteReadOnly);
xe::memory::AllocFixed(generated_code_write_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kReadWrite);
}
} while (generated_code_commit_mark_.compare_exchange_weak(
old_commit_mark, new_commit_mark));
// Copy code.
std::memcpy(code_write_address, machine_code, func_info.code_size.total);
// Fill unused slots with 0xCC
std::memset(tail_write_address, 0xCC,
static_cast<size_t>(end_write_address - tail_write_address));
// Notify subclasses of placed code.
PlaceCode(guest_address, machine_code, func_info, code_execute_address,
unwind_reservation);
}
void X64CodeCache::OnCodePlaced(uint32_t guest_address,
GuestFunction* function_info,
void* code_execute_address, size_t code_size) {
#if ENABLE_VTUNE
if (iJIT_IsProfilingActive() == iJIT_SAMPLING_ON) {
std::string method_name;
if (function_info && function_info->name().size() != 0) {
method_name = function_info->name();
} else {
method_name = xe::format_string("sub_%.8X", guest_address);
method_name = fmt::format("sub_{:08X}", guest_address);
}
iJIT_Method_Load_V2 method = {0};
@@ -278,88 +59,6 @@ void X64CodeCache::PlaceGuestCode(uint32_t guest_address, void* machine_code,
iJIT_NotifyEvent(iJVM_EVENT_TYPE_METHOD_LOAD_FINISHED_V2, (void*)&method);
}
#endif
// Now that everything is ready, fix up the indirection table.
// Note that we do support code that doesn't have an indirection fixup, so
// ignore those when we see them.
if (guest_address && indirection_table_base_) {
uint32_t* indirection_slot = reinterpret_cast<uint32_t*>(
indirection_table_base_ + (guest_address - kIndirectionTableBase));
*indirection_slot =
uint32_t(reinterpret_cast<uint64_t>(code_execute_address));
}
}
uint32_t X64CodeCache::PlaceData(const void* data, size_t length) {
// Hold a lock while we bump the pointers up.
size_t high_mark;
uint8_t* data_address = nullptr;
{
auto global_lock = global_critical_region_.Acquire();
// Reserve code.
// Always move the code to land on 16b alignment.
data_address = generated_code_write_base_ + generated_code_offset_;
generated_code_offset_ += xe::round_up(length, 16);
high_mark = generated_code_offset_;
}
// If we are going above the high water mark of committed memory, commit some
// more. It's ok if multiple threads do this, as redundant commits aren't
// harmful.
size_t old_commit_mark, new_commit_mark;
do {
old_commit_mark = generated_code_commit_mark_;
if (high_mark <= old_commit_mark) break;
new_commit_mark = old_commit_mark + 16_MiB;
if (generated_code_execute_base_ == generated_code_write_base_) {
xe::memory::AllocFixed(generated_code_execute_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kExecuteReadWrite);
} else {
xe::memory::AllocFixed(generated_code_execute_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kExecuteReadOnly);
xe::memory::AllocFixed(generated_code_write_base_, new_commit_mark,
xe::memory::AllocationType::kCommit,
xe::memory::PageAccess::kReadWrite);
}
} while (generated_code_commit_mark_.compare_exchange_weak(old_commit_mark,
new_commit_mark));
// Copy code.
std::memcpy(data_address, data, length);
return uint32_t(uintptr_t(data_address));
}
GuestFunction* X64CodeCache::LookupFunction(uint64_t host_pc) {
uint32_t key = uint32_t(host_pc - kGeneratedCodeExecuteBase);
void* fn_entry = std::bsearch(
&key, generated_code_map_.data(), generated_code_map_.size(),
sizeof(std::pair<uint32_t, Function*>),
[](const void* key_ptr, const void* element_ptr) {
auto key = *reinterpret_cast<const uint32_t*>(key_ptr);
auto element =
reinterpret_cast<const std::pair<uint64_t, GuestFunction*>*>(
element_ptr);
if (key < (element->first >> 32)) {
return -1;
} else if (key > uint32_t(element->first)) {
return 1;
} else {
return 0;
}
});
if (fn_entry) {
return reinterpret_cast<const std::pair<uint64_t, GuestFunction*>*>(
fn_entry)
->second;
} else {
return nullptr;
}
}
} // namespace x64